AI health prediction tools lag in low-income countries, review finds

Artificial intelligence models that promise to forecast disease outbreaks, mortality risk and treatment failure are flourishing in the world’s richest health systems, but a new systematic review finds that the evidence base for using them in low- and middle-income countries remains strikingly thin. Of 64 eligible studies published between January 2018 and March 2025, only 12, or 18.8 percent, were conducted exclusively in low- and middle-income settings, while 44, or 68.8 percent, came from high-income countries and 8 drew on mixed or multi-setting data. The review, published in Discover Artificial Intelligence, followed PRISMA guidelines and searched five major databases, screening 1,248 records down to a final synthesis of 64 peer-reviewed studies that applied machine learning to electronic health records or routinely collected electronic clinical data for public health prediction tasks.

The authors, led by Joe Phiri, set out to answer three questions: which algorithms are most frequently applied in low- and middle-income settings, which preprocessing strategies best address the region’s notorious data quality problems, and how predictions are actually evaluated and validated. Because the direct evidence from resource-constrained settings is still small, high-income studies were deliberately included only as methodological benchmarks, not as proof that methods transfer. That distinction runs through the entire analysis, and the reviewers are explicit that their conclusions about high-income practice cannot simply be read across to health systems with different data completeness, interoperability and infrastructure.

The pattern of model choice splits cleanly along income lines. Across all 64 studies, deep learning architectures, including recurrent neural networks, LSTM models and convolutional neural networks, were the most frequently applied category at 39.1 percent, or 25 studies, driven overwhelmingly by high-income teams with access to large, curated datasets. Among the 12 studies focused exclusively on low- and middle-income settings, however, traditional machine learning and ensemble methods were each the dominant approach, applied in 33.3 percent of studies apiece. Logistic regression and generalised linear models accounted for 28.1 percent of all studies, ensembles such as random forests and gradient boosting variants for 20.3 percent, and transformer-based models like Med-BERT for 6.2 percent, all of the latter conducted in high-income settings.

The reviewers argue that the low- and middle-income distribution should be read as practical alignment rather than methodological lag. Traditional and ensemble models are interpretable, computationally light and calibrate more reliably under data constraints, which matters where electricity, internet connectivity and computing hardware are unreliable. Their position is supported by prior comparative evidence showing that deep learning offers modest or no performance advantage over logistic regression under realistic data conditions. In other words, algorithmic sophistication alone does not guarantee better public health utility when the underlying data are sparse or incomplete.

The data feeding these models differ profoundly between settings. Many primary care facilities, particularly in sub-Saharan Africa, still rely on paper or hybrid records, and where electronic systems exist they are typically platforms such as OpenMRS, DHIS2 or Zambia’s SmartCare, designed for programme reporting rather than research-grade longitudinal data. The review catalogues the resulting problems: systematic missingness as patients seek care at multiple unlinked facilities, inconsistent coding of diagnoses and laboratory results, limited longitudinal depth due to patient mobility, and incomplete digitisation of historical records. Data completeness in the low- and middle-income studies frequently fell below 80 percent for key clinical variables, in contrast with the more standardised multi-institutional datasets common in high-income systems built on platforms such as Epic and Cerner.

Preprocessing practice reflected these constraints. Most studies reported basic imputation, typically mean or median substitution, for missing data, while multiple imputation appeared mainly in high-income work. Feature engineering was usually based on clinically defined variables rather than automated representation learning, which was rarely feasible in low- and middle-income research because of data sparsity and computational limits. The reviewers identify a critical gap here: almost no studies performed sensitivity analyses showing how preprocessing choices propagate to downstream predictions, meaning reported performance cannot be clearly attributed to model architecture rather than data preparation decisions.

The most alarming findings concern evaluation. Discrimination metrics, especially the area under the receiver operating characteristic curve, dominated reporting, with sensitivity and specificity also common. Yet calibration, the property that predicted probabilities match observed outcomes and which is essential when predictions determine triage thresholds, was reported in only 4 of 64 studies, or 6.2 percent. External validation on an independent dataset appeared in just 5 studies, or 7.8 percent, comprising 3 high-income studies, 1 low- and middle-income study and 1 mixed-setting study. Nearly every study relied on internal validation through cross-validation or train-test splits, which the reviewers say substantially limits confidence in the generalisability of reported performance across contexts.

Explainability reporting showed an equally stark asymmetry. Overall, 17 of 64 studies, or 26.6 percent, included an explainability assessment, most commonly through SHAP-based feature importance rankings or post-hoc attribution techniques. But the split was 16 of 44 high-income studies, or 36.4 percent, against just 1 of 12 low- and middle-income studies, or 8.3 percent. The reviewers frame this not as a reporting omission but as a deployment-readiness gap: where model outputs influence triage, follow-up or service prioritisation under conditions of scarcity, explainability is tied to accountability, safe override of unsafe outputs and user trust, and it is least present precisely where institutional oversight of algorithms is weakest and the consequences of unexplained errors fall on underserved populations.

The disease focus of the low- and middle-income literature also lags the epidemiological reality. Prediction targets clustered around HIV, tuberculosis and other infectious diseases at 45.3 percent, non-communicable diseases at 28.1 percent, maternal and child health at 14.1 percent and health system utilisation at 12.5 percent. While the infectious disease emphasis aligns with leading causes of premature death in sub-Saharan Africa, conditions such as cardiovascular disease, diabetes and hypertension are underrepresented relative to their rapidly growing burden, suggesting the research still mirrors the structure of historically digitised programmes rather than the current demand for scalable chronic disease prediction.

The review has limitations the authors acknowledge: the protocol was not prospectively registered, the search was restricted to English-language publications, only a minority of studies provided direct low- and middle-income evidence, and heterogeneity across outcomes and datasets precluded quantitative meta-analysis. Still, the priorities they draw are concrete. Model selection should default to robustness over complexity, preprocessing decisions should be tested through sensitivity analyses, calibration and external validation should be treated as core requirements rather than optional extras, and explainability should be evaluated for whether it actually improves clinician understanding and supports safe disagreement with model outputs. Future work should also expand non-communicable disease applications and follow emerging reporting standards such as TRIPOD+AI and MI-CLAIM. The central diagnosis is a structural imbalance: investment in algorithmic development has outpaced investment in evaluation rigour, explainability and governance, and the countries bearing the greatest burden of preventable disease are precisely where that imbalance is most acute.

Subject of Research: Machine learning prediction using electronic health records for public health in low- and middle-income countries

Article Title: A systematic review of artificial intelligence and machine learning for public health predictions using electronic health records in low and middle income countries with comparative benchmarks from high income settings

Article References: Phiri, J., Zimba, A., Njovu, C., & Lumpa, M. (2026). A systematic review of artificial intelligence and machine learning for public health predictions using electronic health records in low and middle income countries with comparative benchmarks from high income settings. Discover Artificial Intelligence, 6(1), Article 944. https://doi.org/10.1007/s44163-026-01958-x

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01958-x

Keywords: artificial intelligence, machine learning, electronic health records, public health, low- and middle-income countries, systematic review, model calibration, external validation, explainable AI, deep learning, ensemble methods, health data quality